// RUN: tf-tfrt-opt -verify-diagnostics -split-input-file -tfrt-fuse-tpu-compile-and-execute-ops -canonicalize %s | FileCheck %s --dump-input=fail --dump-input-filter=all

module attributes {tf_saved_model.semantics} {

// Test fusing _TPUCompileMlirOp and TPUExecuteOp into TPUCompileMlirAndExecuteOp.

// CHECK-LABEL: func private @test_fuse_tpu_ops
func.func private @test_fuse_tpu_ops(%arg0: tensor<*xi32>, %arg1: tensor<*x!tf_type.resource>) -> tensor<*xi32> {

     // var <- var - mom
     (TF_AssignSubVariableOp $var_resource, $mom_new)
   ]
   >;

// Same as DecomposeResourceApplyFtrlV2, with l2_shrinkage set to zero.
// TODO(kramm): Move this pattern to canonicalize.td?
def DecomposeResourceApplyFtrl : Pat<
  (TF_ResourceApplyFtrlOp $var, $accum, $linear, $grad, $lr, $l1, $l2,
     $lr_power, $use_locking, $multiply_linear_by_lr),

// RUN: tfr-opt %s -tfr-decompose -tfr-raise-to-tf -canonicalize -verify-diagnostics -split-input-file | FileCheck %s

//=================> User models, from GraphDef <====================

// CHECK-LABEL: my_identity
func.func @my_identity(%arg0: tensor<2x3xf32>) -> tensor<2x3xf32> {
  %0 = "tf.MyIdentity"(%arg0) : (tensor<2x3xf32>) -> tensor<2x3xf32>
  func.return %0 : tensor<2x3xf32>

// CHECK-NEXT: return %arg0 : tensor<2x3xf32>
}

      point = point < 0 ? dim_i + point : point;
      return Clamp(point, bounds[0], bounds[1]);
    };

    begin_i = canonicalize(begin_i, 0);
    end_i = canonicalize(end_i, 1);

    int64_t interval_len = end_i - begin_i;
    int64_t size_i = 0;
    // If internal length is zero or has different sign from stride, it's a
    // degenerated case: we are slicing nothing. Otherwise, calculate the sliced

	}
	warnings = append(warnings, podutil.GetWarningsForPod(ctx, newPod, nil)...)
	return warnings
}

// Canonicalize normalizes the object after validation.
func (podStrategy) Canonicalize(obj runtime.Object) {
}

// AllowCreateOnUpdate is false for pods.
func (podStrategy) AllowCreateOnUpdate() bool {
	return false
}

func SetDefaults_Namespace(obj *v1.Namespace) {
	// we can't SetDefaults for nameless namespaces (generateName).
	// This code needs to be kept in sync with the implementation that exists
	// in Namespace Canonicalize strategy (pkg/registry/core/namespace)

	// note that this can result in many calls to feature enablement in some cases, but
	// we assume that there's no real cost there.
	if len(obj.Name) > 0 {

	if v == nil {
		bugLogIf(ctx, errInvalidArgument)
		return errInvalidArgument
	}

	nv := make(textproto.MIMEHeader, len(v))
	for k, kv := range v {
		// Canonicalize all headers, to remove any duplicates.
		nv[http.CanonicalHeaderKey(k)] = kv
	}

	// Save all supported headers.
	for _, supportedHeader := range supportedHeaders {

	}
	return getLangISO3(s)
}

// TODO language normalization as well as the AliasMaps could be moved to the
// higher level package, but it is a bit tricky to separate the generation.

func (id Language) Canonicalize() (Language, AliasType) {
	return normLang(id)
}

// normLang returns the mapped langID of id according to mapping m.
func normLang(id Language) (Language, AliasType) {

      expand_axis = (*mlir::cast<DenseElementsAttr>(const_op.getValue())
                          .getValues<APInt>()
                          .begin())
                        .getSExtValue();
      // Canonicalize axis. Some TF python functions, such as
      // `tf.nn.convolution`, use negative axis.
      if (expand_axis < 0) {
        // Always expand 3D input to 4D input.
        expand_axis += 4;
      }
    } else {

	NamedValues map[LocalSlot][]*Value
	// Names is a copy of NamedValues.Keys. We keep a separate list
	// of keys to make iteration order deterministic.
	Names []*LocalSlot
	// Canonicalize root/top-level local slots, and canonicalize their pieces.
	// Because LocalSlot pieces refer to their parents with a pointer, this ensures that equivalent slots really are equal.
	CanonicalLocalSlots  map[LocalSlot]*LocalSlot